Unexpected issues with xtdidregress

[Geoff Zheng]

I wanted to ensure that xtdidregress was consistent with the "old-fashioned" way of running differences in differences. To do this, I generated a simulated dataset and ran some regressions (see code).

I started by verifying that `reg` and `reghdfe` produced consistent results. To my surprise, in this toy dataset, xtdidregress wouldn't run due to collinearity, even when I included the `nogteffects` option (so that there should be nothing to be co-linear with)!

What am I doing wrong?

clear *
set seed 1234
local N 50

set obs `N'
gen id = _n

gen x = rnormal()
expand 2, gen(t)
label val t

* y is independent of x for t = 0
* y is related to x at t = 1
gen y = rnormal() + (t==1) * x

xtset id t

estimates clear
eststo: reghdfe y c.x##i.t, a(t id)

eststo: reg D.y x

esttab *, keep(1.t#c.x x)

* STATA complains about collinearity
xtdidregress (y) (x, continuous), group(id) time(t) nogteffects

/*
model is not identified
    The treatment variable x was omitted because of collinearity.

*/

* Issue isn't with the continuous treatment
egen x_bin = cut(x), group(2)
xtdidregress (y) (x_bin), group(id) time(t) nogteffects

[Enrique Pinzon (StataCorp)]

Dear Geoff,

What goes in the second parenthesis should be the observation level treatment variable. In your case, -1.t#c.x-. Try for example:

generate treat = 1.t#c.x
xtdidregress (y) (treat, continuous), group(id) time(t)

This should give you what you are expecting.

-xtdidregress- runs -xtreg..., fe-. When -xtreg, fe- drops the observation level treatment from the model, the command refuses to proceed. In your case, with -nogteffects-:

. xtreg y x, fe vce(cluster id)
note: x omitted because of collinearity.

Fixed-effects (within) regression               Number of obs     =        100
Group variable: id                              Number of groups  =         50

R-squared:                                      Obs per group:
     Within  =      .                                         min =          2
     Between =      .                                         avg =        2.0
     Overall =      .                                         max =          2

                                                F(0, 49)          =          .
corr(u_i, Xb) =      .                          Prob > F          =          .

                                    (Std. err. adjusted for 50 clusters in id)
------------------------------------------------------------------------------
             |               Robust
           y | Coefficient  std. err.      t    P>|t|     [95% conf. interval]
-------------+----------------------------------------------------------------
           x |          0  (omitted)
       _cons |  -.0147431          .        .       .            .           .
-------------+----------------------------------------------------------------
     sigma_u |  .79912424
     sigma_e |  1.2827061
         rho |  .27960503   (fraction of variance due to u_i)
------------------------------------------------------------------------------

[Geoff Zheng]

Thanks for the clarification. This was very helpful